DQN-Based Channel-Adaptive Federated Learning over Unreliable Wireless Channels with Hybrid Gradient Compression

Abstract

Federated learning (FL) over wireless networks suffers from two compounding degradations: stochastic channel fading that induces gradient transmission failures, and fixed-rate compression policies that either waste bandwidth under good channel conditions or sacrifice model fidelity under poor ones. This paper proposes a Deep Q-Network (DQN)-based channel-adaptive FL framework that dynamically selects the optimal gradient compression policy jointly controlling top-k sparsification ratio and quantization bit-width—for each participating client in every communication round, conditioned on real-time channel state information (CSI). The DQN agent observes a three-dimensional state comprising instantaneous SNR, achievable data rate, and local data volume, and outputs one of four discrete compression actions ranging from full-precision 16-bit transmission to aggressive 4-bit quantization with 90% sparsification. A wireless channel simulator models Rayleigh fading, bandwidth-limited Shannon capacity, and outage events, providing a realistic training environment for the agent. Experiments on the Fashion-MNIST benchmark with non-IID client data partitioning demonstrate that the proposed method achieves 93.18% test accuracy a gain of 4.37–8.86 percentage points over three strong baselines—while reducing total transmission energy by 55.6% and communication overhead by 58.9% compared to FedAvg without compression. The DQN agent converges to a coherent channel-aware policy: selecting high-quality (low-compression) actions for clients with strong channel conditions and aggressive compression for outage-prone clients, thereby maximizing successful participation. These results establish the proposed framework as a principled, resource-efficient solution for deploying FL in practical wireless communication systems.